ENPIRE: eight agents, real robots, and who pays for each attempt
NVIDIA's ENPIRE scaled from one Codex agent to eight on physical robot tasks, cutting pin-insertion time from 1.5 hours to 40 minutes, but token costs grew superlinearly and two of three agents failed Push-T on real hardware.
Eight Codex agents, eight stations with YAM bimanual robots, an RTX 5090 on each. The project is ENPIRE, from NVIDIA, CMU, and UC Berkeley. Each agent gets a robot and a physical task: inserting pins into holes, handling zip ties, plugging a GPU into a slot. The agent writes code, runs the robot, checks the outcome, revises the control policy, and tries again.
Scaling speeds up the result. On pin insertion, a single agent needed over 1.5 hours to reach near-100% success. Eight agents: about 40 minutes. On Push-T, time to a normalized score of 1.0 dropped from roughly five hours to two.
Two out of three agents failed Push-T on the real robot, though every agent solved it in simulation. The robots and GPUs sit idle while the agent reads logs, writes code, and waits for the model. Scaling from one agent to eight, token consumption grows faster than linearly: you gain speed, but coordination overhead grows too.
Each physical attempt takes time: resetting the scene, running the robot, checking the result. Even so, NVIDIA reports 99% pass@8 (at least one success in eight attempts) on the showcased tasks.
If systems like this become reliable, control shifts to whoever sets the goal, writes the constraints, checks the result, and pays for physical mistakes.
Experiment is an organization on Eternal Search.
Open the related Eternal Search page
Sources
research.nvidia.com
t.me
t.me
ENPIRE is selected because it makes a structural claim visible with measured data: when AI absorbs hypothesis generation and policy iteration, the bottleneck shifts to the cost per physical attempt. Experiment on Eternal Search is an organization that funds early-stage scientific research, precisely the layer the ENPIRE authors themselves identify as decisive in their own conclusion — those who pay for physical mistakes hold control. The connection is structural, not a direct funding relationship.